The GEO paper counted citations inside sourced answers

You keep seeing the same advice: optimize for AI search, do GEO, get cited by ChatGPT. Most of it is recycled blog posts selling tools.

There is a real research paper under all that noise. In 2024, researchers from Princeton, Georgia Tech, and IIT Delhi published GEO: Generative Engine Optimization (arXiv:2311.09735). They worked out what “visibility” even means when an assistant answers in paragraphs with inline citations, and then they tested specific content changes to see which ones helped.

You don’t need to read the whole paper. This article pulls out the parts that matter if you run a site. GEO measured which pages get quoted inside a sourced answer. The test never checked which products or places get recommended.

1. GEO, in one plain sentence

GEO, or Generative Engine Optimization, means changing how your page is written and structured so that a generative engine building an answer from sources is more likely to use your page, cite it, and give it weight. By generative engine, the paper means Perplexity-style search, Bing Chat–style answers, Google AI Overviews–style summaries, and similar systems.

That differs from classic SEO. SEO mostly competes for a position in a ranked list of blue links, while GEO competes for a share of one synthesized answer: which sentences come from you, where they sit, and whether the system attributes the claim to you.

The four big assistants, ChatGPT, Claude, Perplexity, and Gemini, don’t retrieve the same way, and the paper’s setup is closest to search, then summarize with citations. Treat what follows as a solid pattern for that class of tool, not as a ranking algorithm for every assistant.

2. Visibility is not “rank #3” anymore

The paper puts the same kind of page side by side, once as a soft claim and once as a claim a reader can check.

In classic Google results, rank is mostly a position on a list. In a generative answer the sources are woven into one block, so one site might supply the opening paragraph, another might get a brief mention in the middle, and a third might never be named even though it ranks well in ordinary search.

That is why the paper argues for different ways of thinking about an impression:

You don’t need their formulas. The job is to be a source the model can safely lean on, with clear claims, checkable detail, and sentences that are easy to quote. That is a different job from repeating a query phrase until Google notices.

3. What the paper actually tested (and what worked)

The researchers took pages that were already in a top-five retrieval set, applied one content change at a time, and measured whether that page’s impression in the generated answer went up.

They tried nine tactics:

• Adding statistics (numbers instead of vague adjectives)
• Adding quotations from credible sources
• Citing sources on the page itself
• Easier wording
• Smoother sentences
• Authoritative tone
• Technical terms
• Unique words
• Classical keyword stuffing

What moved the needle

Their strongest results clustered around credibility and concrete detail: statistics, quotations, and citations on the page. For the best methods, they report relative visibility lifts of roughly 30–40% on their main objective metric, with solid gains on subjective impression too. Clearer, more fluent writing also helped.

Keyword stuffing, the classic SEO reflex, did poorly. In their numbers it often failed to help and sometimes came out worse than doing nothing, which is worth remembering if someone is still selling you “more AI keywords.”

An authoritative or persuasive tone on its own was not a winner either. Sounding confident is cheaper than being checkable, and generative engines already see plenty of marketing voice.

What that looks like in practice

Statistics (a how-to page).
Weak: “This hotel lobby is perfect for remote work.”
Stronger: “Twelve two-person tables, power at every seat, Wi‑Fi ~80 Mbps down on a weekday afternoon test — doors open 7:30–18:00 Mon–Fri.”

Numbers aren’t decoration here. They are the kind of sentence an answer can repeat without inventing anything.

Quotations (a place or experience page).
Weak: “Locals love this walking tour in spring.”
Stronger: quote the municipal tourism page or the park authority on the season, access, or closures, and keep your own tips separate from that.

A model summarizing “is this walking tour running with kids in March?” then has something grounded to lift.

Cite sources (a deal-check or explainer).
Weak: “Experts say this vintage is underpriced.”
Stronger: “Producer tech sheet lists 13.5% ABV and 18 months in oak; Wine-Searcher average for this vintage in our market was €X last month,” with the links.

None of that is an SEO hack. You are giving the answering system a trail it can attribute.

4. Don’t overread the “+40%”

This section exists so that GEO blogs can’t sell you a fantasy. The paper is useful; the marketing around it often isn’t. Three misreadings to avoid:

1. GEO is not step one.
In the study, the page was already among the sources the engine had fetched, and the researchers then made that page more citable. If crawlers can’t reach you, or you’re behind a login, or your facts contradict each other, rewriting copy for GEO won’t rescue you. Fix fetching and facts first (Your catalog can be perfect and still never show up).

2. “+40%” does not mean 40% more customers.
That figure describes a larger share of the answer among pages that were already competing inside the same response. It means more of the generated text leaning on you, not traffic, sales, or ChatGPT naming your product. The same goes for their Perplexity test at around 37%: promising for citation share, not a universal ranking score.

3. Real numbers only.
Statistics and quotes helped because they were genuine content upgrades. Invented percentages, fake studies, and AI-generated “expert” quotes are not GEO, and they are how you end up misstated or ignored later. If you don’t have a figure you can defend, don’t invent one. If the page shows an updated date, the figures under it have to be from that date — a fresh stamp on last year’s number makes a stale claim look current.

One more caveat: what works on a factual how-to may not work the same way on a local place page or a debate-style question. Take the idea, which is to be specific and citeable, rather than one rigid template for every URL.

What to do with this (without becoming a GEO agency)

1. Pick one important page that already ranks or gets cited occasionally, such as a guide, a place page, or a deal explainer, rather than your thinnest product page.
2. Add one real statistic or measurable detail where it belongs.
3. Add one real quote or primary citation, with a link.
4. Cut the keyword stuffing and the empty superlatives.
5. Leave schema, feeds, and robots.txt to their own jobs, and don’t confuse them with GEO copy work.
6. When a niche matters to you, run your own prompts later. Don’t claim you “tested across models” until you have.

Four lines are worth keeping from the research: GEO is about share of the answer rather than blue-link rank, statistics and quotes and citations helped, keyword stuffing didn’t, and none of it matters if you were never retrieved in the first place.

Being cited in a sourced answer is still not the same as being named on a ChatGPT shortlist — Your page can be used as a source without you being recommended

Source: Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024 — arXiv:2311.09735. Also listed on Resources.

Read next if: Your page can be used as a source without you being recommended · Those AI visibility tools just count mentions